ai-benchmarking

Suicide Risk Assessment AI Benchmark

This repository provides a standardized framework for benchmarking Large Language Models (LLMs) against the Columbia-Suicide Severity Rating Scale (C-SSRS). It evaluates AI models on their ability to accurately categorize risk and provide safe, urgent crisis instructions.

---

1. Quickstart

⚙️ Setup

1. Environment: Ensure you have uv installed.

2. Dependencies: Install the project environment:

   uv sync
   

3. API Keys: Create a private .env file in the ai_benchmarking/ directory. Do not edit .env.example with real keys.

   OPENAI_API_KEY=your_key GEMINI_API_KEY=your_key_here ANTHROPIC_API_KEY=your_key_here

🚀 Running the Benchmark

You can evaluate different models by changing the --provider and --model flags. Use a fast, low-cost model as the --judge-model to save on API costs.

1. Google Gemini (Recommended)

   The Gemini 3 series is highly efficient for both inference and judging. For model names, see https://docs.cloud.google.com/gemini-enterprise-agent-platform/resources/locations#google-models

   • Inference Model: gemini-3.1-flash-lite (Fastest/Cheapest) or gemini-3.1-pro-preview (High Reasoning)

   • Judge Model: gemini-3.5-flash

   uv run python -m ai_benchmarking.eval \
     --provider gemini \
     --model gemini-3.1-flash-lite \
     --data data/input.json \
     --output outputs/gemini_results.json \
     --kb data/knowledge_base.json \
     --judge-model gemini-3.5-flash
   

2. Anthropic Claude The Claude 4 series provides industry-leading clinical nuance. For model names, see https://platform.claude.com/docs/en/about-claude/models/overview

   • Inference Model: claude-4-sonnet-20260217 or claude-4-haiku-20251015

   • Judge Model: claude-4-sonnet-20260217

   uv run python -m ai_benchmarking.eval \
     --provider anthropic \
     --model claude-4-sonnet-20260217 \
     --data data/input.json \
     --output outputs/claude_results.json \
     --kb data/knowledge_base.json \
     --judge-model gemini-3.1-flash-lite
   

3. OpenAI OpenAI's latest "O-series" models are built for deep reasoning and safety. For model names, see https://platform.openai.com/chat/edit.

   • Inference Model: gpt-5.2-chat-latest or o5-mini

   • Judge Model: gpt-5.1-mini

   uv run python -m ai_benchmarking.eval \
     --provider openai \
     --model o5-mini \
     --data data/input.json \
     --output outputs/openai_results.json \ 
     --kb data/knowledge_base.json \
     --judge-model gemini-3.1-flash-lite
   

You may also evaluate a --model against a --data dataset without using --kb to augment the base prompt, and without using --judge-model to skip evaluating safety and communication_quality dimensions. This is particularly useful to test a new model's ability to accurately classify with just our zero-shot base_system_prompt against the human-evaluated knowledge_base.json, aka our "Gold Master" dataset, by running:

```bash
MODEL=gemini-3.1-flash-lite
uv run python -m ai_benchmarking.eval \
  --provider gemini \
  --model $MODEL \
  --data data/knowledge_base.json \
  --output outputs/kb_baseline.${MODEL}.json
```

Normally, the input --data is assumed to be predicted scores from a lighterweight setup (such as output by process_batches.py), that are to be AI-evaluated using our kb-augmented model. The above command, however, switches the roles that the --data and --model play while still testing for label equality to get accuracy, cost and latency metrics.

🛠 Model Selection Guide (May 2026)

• Lowest Cost: gemini-3.1-flash-lite (Provider: Google)
• High Precision: claude-4-sonnet-20260217 (Provider: Anthropic)
• Deep Reasoning: gpt-5.2-chat-latest (Provider: OpenAI)
• Best Judge: gemini-3-flash (Provider: Google)

---

2. The Risk Framework (C-SSRS)

Each input is categorized into one of nine labels based on clinical thresholds:

3. Code Architecture

• ai_benchmarking/inference.py: The "Router." Handles API calls to OpenAI, Google (via modern google-genai SDK), and Anthropic. It enforces a strict JSON output format.
• ai_benchmarking/eval.py: The "Engine." Iterates through the dataset, triggers inference, calls the judge, and maps numeric scores to clinical labels for accuracy checking.
• ai_benchmarking/judges.py: The "Auditor." An LLM-as-a-Judge that reviews the AI's response for Safety (escalation for codes 6-8) and Communication Quality.
• ai_benchmarking/utils.py: The "Accountant." Calculates real-time costs based on 2026 token pricing and computes aggregate accuracy, latency, and cost statistics.

---

4. Metrics Tracked

• Accuracy: % of AI-predicted labels that exactly match the expert Ground Truth labels.
• Safety Pass Rate: % of responses that met emergency protocol requirements for high-risk queries.
• Latency: Round-trip time in seconds (crucial for time-sensitive crisis intervention).
• Cost: Calculated using provider-specific pricing per 1 million tokens (May 2026 rates).

---

⚖️ License

This project is licensed under the GNU GPL v3. We chose this license to ensure that improvements to this suicide risk benchmarking logic remain open and accessible to the entire non-profit and mental health community.